OpenAI trialling text watermarks for ChatGPT and Codex in EU
OpenAI is introducing an invisible, machine-readable watermark for text generated by ChatGPT and Codex, initially for eligible users in the European Union. The move is intended to help address transparency requirements under the EU’s AI Act, while giving OpenAI a chance to assess the system through real-world use and feedback.
The rollout will begin over the coming weeks and cover all plans in the EU; watermarking will not be a global default at launch. API customers worldwide can opt in for selected models, and cloud partners are expected to offer it in the coming weeks. OpenAI says its textGrain system performed comparably to unwatermarked text on cited benchmarks, but warns it cannot reliably detect all watermarked text or establish accuracy, ownership, human contribution or authorship. Approved researchers and expert organisations may apply for detector access, which will be granted case by case.
- ChatGPT and Codex text watermarking is rolling out in the EU.
- API customers worldwide can opt in for selected models.
- OpenAI says watermarks cannot prove accuracy or human authorship.
New here? Start with this
ChatGPT and Codex are artificial intelligence tools created by OpenAI that can generate text and computer code. As AI-generated content becomes more common, there is growing concern about whether people can tell machine-written text from human-written text. Watermarks are a potential solution: invisible markers that machines can read to identify AI-generated content.
The European Union has introduced new regulations called the AI Act to govern how artificial intelligence is used and deployed. These rules require companies to be transparent about when they use AI and to disclose which content has been created by machines. OpenAI has developed a watermarking system to help meet these transparency requirements and to test how the technology performs in real-world use.
OpenAI's watermarking system embeds invisible markers into text generated by its tools, without changing how the text reads or functions. However, the watermarks have limitations: they cannot always be detected, the system cannot distinguish between human-written and machine-generated text, and it cannot determine what proportion of text was created by humans versus machines.
Both sides, in good faith
The strongest fair case each way — we don't pick a winner.
The case for
Transparency about AI-generated content serves legitimate public interests in informed decision-making and preventing fraud and misinformation. Invisible watermarks provide technical infrastructure for the EU AI Act's disclosure requirements, and even imperfect systems offer meaningful value in establishing accountability norms around AI deployment. OpenAI's trial approach, with opt-in choices for API users, represents good-faith effort to balance regulatory intent with legitimate variation in use cases.
The case against
The watermarking system's limitations—OpenAI's own admission that it cannot reliably detect watermarks or establish authorship—fundamentally undermine the technology's core purpose as a transparency tool. Watermarks can be circumvented through text transformation, creating dangerous false confidence in detection capabilities without providing genuine security. The fragmented rollout across regions and user types lacks the scale needed for meaningful transparency impact, and resources might be better directed toward clearer labelling systems that actually inform users and regulators.
Read the full article at the source →
Originally published by The Verge as “OpenAI is adding text watermarking in ChatGPT and Codex”.